122 lines · plain
1// RUN: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=lane" \2// RUN: | mlir-runner \3// RUN: --shared-libs=%mlir_levelzero_runtime \4// RUN: --shared-libs=%mlir_runner_utils \5// RUN: --entry-point-result=void \6// RUN: | FileCheck %s7 8module @gemm attributes {gpu.container_module} {9 gpu.module @kernel {10 gpu.func @simple_gemm(%a: memref<256x256xf16>, %b: memref<256x256xf16>, %c: memref<256x256xf32>) kernel {11 %c0 = arith.constant 0 : index12 %c1 = arith.constant 1 : index13 %c8 = arith.constant 8 : index14 %c16 = arith.constant 16 : index15 %c32 = arith.constant 32 : index16 %c256 = arith.constant 256 : index17 %block_x = gpu.block_id x18 %block_y = gpu.block_id y19 %x_block_offset = arith.muli %block_x, %c8 : index20 %y_block_offset = arith.muli %block_y, %c16 : index21 22 %c_tdesc = xegpu.create_nd_tdesc %c : memref<256x256xf32> -> !xegpu.tensor_desc<8x16xf32>23 %c_init_value = xegpu.load_nd %c_tdesc[%x_block_offset, %y_block_offset] : !xegpu.tensor_desc<8x16xf32> -> vector<8xf32>24 %a_tdesc = xegpu.create_nd_tdesc %a : memref<256x256xf16> -> !xegpu.tensor_desc<8x16xf16>25 %b_tdesc = xegpu.create_nd_tdesc %b : memref<256x256xf16> -> !xegpu.tensor_desc<16x16xf16>26 27 %r = scf.for %k = %c0 to %c256 step %c16 iter_args(%arg_c = %c_init_value) -> (vector<8xf32>) {28 %a_val = xegpu.load_nd %a_tdesc[%x_block_offset, %k] : !xegpu.tensor_desc<8x16xf16> -> vector<8xf16>29 %b_val = xegpu.load_nd %b_tdesc[%k, %y_block_offset] : !xegpu.tensor_desc<16x16xf16> -> vector<16xf16>30 %dpas = xegpu.dpas %a_val, %b_val, %arg_c : vector<8xf16>, vector<16xf16>, vector<8xf32> -> vector<8xf32>31 scf.yield %dpas : vector<8xf32>32 }33 xegpu.store_nd %r, %c_tdesc[%x_block_offset, %y_block_offset] <{l1_hint = #xegpu.cache_hint<write_back>, l2_hint = #xegpu.cache_hint<uncached>}>: vector<8xf32>, !xegpu.tensor_desc<8x16xf32>34 gpu.return35 }36 }37 38 func.func @test(%a : memref<256x256xf16>, %b : memref<256x256xf16>, %c : memref<256x256xf32>) -> memref<256x256xf32> attributes {llvm.emit_c_interface} {39 %c1 = arith.constant 1 : index40 %c16 = arith.constant 16 : index41 %c32 = arith.constant 32 : index42 %memref_a = gpu.alloc () : memref<256x256xf16>43 gpu.memcpy %memref_a, %a : memref<256x256xf16>, memref<256x256xf16>44 %memref_b = gpu.alloc () : memref<256x256xf16>45 gpu.memcpy %memref_b, %b : memref<256x256xf16>, memref<256x256xf16>46 %memref_c = gpu.alloc () : memref<256x256xf32>47 gpu.memcpy %memref_c, %c : memref<256x256xf32>, memref<256x256xf32>48 gpu.launch_func @kernel::@simple_gemm blocks in (%c32, %c16, %c1) threads in (%c16, %c1, %c1) args(%memref_a : memref<256x256xf16>, %memref_b : memref<256x256xf16>, %memref_c : memref<256x256xf32>)49 gpu.wait // Wait for the kernel to finish.50 gpu.memcpy %c, %memref_c : memref<256x256xf32>, memref<256x256xf32>51 gpu.dealloc %memref_a : memref<256x256xf16>52 gpu.dealloc %memref_b : memref<256x256xf16>53 gpu.dealloc %memref_c : memref<256x256xf32>54 return %c : memref<256x256xf32>55 }56 57 func.func @main() attributes {llvm.emit_c_interface} {58 %c0 = arith.constant 0 : index59 %c1 = arith.constant 1 : index60 %c1_f16 = arith.constant 1.0 : f1661 %c2_f16 = arith.constant 2.0 : f1662 %c256 = arith.constant 256 : index63 %cf_0 = arith.constant 0.0 : f1664 %cf_1 = arith.constant 1.0 : f1665 %A = memref.alloc() : memref<256x256xf16>66 %B = memref.alloc() : memref<256x256xf16>67 %C = memref.alloc() : memref<256x256xf32>68 %C_ref = memref.alloc() : memref<256x256xf32>69 %c_gen_int = arith.constant 0 : i170 %cf_lower = arith.constant -0.5 : f3271 %cf_upper = arith.constant 0.5 : f3272 73 // Initialize matrix A ; A[i, j] = j74 scf.for %i = %c0 to %c256 step %c1 {75 scf.for %j = %c0 to %c256 step %c1 {76 %t = index.castu %j : index to i1677 %val = arith.uitofp %t : i16 to f1678 memref.store %val, %A[%i, %j] : memref<256x256xf16>79 }80 }81 82 // Initialize the B matrix.83 // Make matrix B an identity matrix.84 scf.for %i = %c0 to %c256 step %c1 {85 scf.for %j = %c0 to %c256 step %c1 {86 %i_i32 = index.castu %i : index to i3287 %j_i32 = index.castu %j : index to i3288 %i_j_same = arith.cmpi eq, %i_i32, %j_i32 : i3289 90 scf.if %i_j_same {91 memref.store %cf_1, %B[%i, %j] : memref<256x256xf16>92 } else {93 memref.store %cf_0, %B[%i, %j] : memref<256x256xf16>94 }95 }96 }97 98 // Initialize matrix C and C_ref ; C[i, j] = 099 %c0_f32 = arith.constant 0.0 : f32100 scf.for %i = %c0 to %c256 step %c1 {101 scf.for %j = %c0 to %c256 step %c1 {102 memref.store %c0_f32, %C[%i, %j] : memref<256x256xf32>103 memref.store %c0_f32, %C_ref[%i, %j] : memref<256x256xf32>104 }105 }106 107 // Run GPU version.108 %2 = call @test(%A, %B, %C) : (memref<256x256xf16>, memref<256x256xf16>, memref<256x256xf32>) -> memref<256x256xf32>109 %gpu_result_cast = memref.cast %2 : memref<256x256xf32> to memref<*xf32>110 111 // CHECK: Unranked Memref base@ = 0x{{[0-9a-f]+}}112 // CHECK-COUNT-256: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]113 call @printMemrefF32(%gpu_result_cast) : (memref<*xf32>) -> ()114 memref.dealloc %A : memref<256x256xf16>115 memref.dealloc %B : memref<256x256xf16>116 memref.dealloc %C : memref<256x256xf32>117 memref.dealloc %C_ref : memref<256x256xf32>118 return119 }120 func.func private @printMemrefF32(memref<*xf32>) attributes {llvm.emit_c_interface}121}122